"Introduction. In the past decade, more and more attention has been paid to collaborative learning. A major theme in the collaborative learning field is why some groups are more successful than others (Barron, 2003; Suthers, 2006). Lately, researchers have sought to address this issue by analyzing interaction processes in collaborative learning reasoning that human cognition is based on interactions between individuals and social context or community (Engeström, 1987). Various methods have been developed in previous research to analyze interactions. The following analytic methods have been widely used: (a) Conversation analysis (Sacks, 1962,1995), identifying closings and openings of action sequences (Zemel, Xhafa, & Stahl, 2005); (b) Social network analysis (Wasserman & Faust, 1994) , investigating patterns of interaction (de Laat, Lally, Lipponen, & Simons, 2007) and examining the response relations among participants during online discussions (Aviv, Erlich, Ravid, & Geva, 2003; De Liddo et al., 2011); (c) Content analysis (Chi, 1997), using coding schemes to categorize and count user actions to analyze argumentative knowledge construction (Weinberger & Fischer, 2006), evidence use for the knowledge building principles (van Aalst & Chan, 2007), depth of understanding (Zhang, Scardamalia, Reeve, & Richard, 2009); (d) Sequential analysis, using transitional state diagrams to compute transitional probabilities between coded discourse moves in argumentation (Jeong et al., 2011). Each method has limitations. Table 1 summarizes the analysis approaches, focus and limitations of the different methods. Table. 1. Comparison of different analysis methods. At present, the most often used method is content analysis (Strijbos & Stahl, 2007). Content analysis technique is defined as “a research methodology that builds on procedures to make valid inferences from text” (Rourke, Anderson, Garrison, & Archer, 2001). The essential step of content analysis is to code discussions according to the selected coding scheme. However, different researchers put forward different coding schemes. Well-known examples include content coding schemes for the analysis of the learning process in computer conferencing (Henri, 1992), co-construction of understanding and knowledge (Zhu, 1996), the social construction of knowledge in computer conferencing (Gunawardena, 1997), the social presence in the community of inquiry (Rourke, 1999), the collaborative construction of knowledge (Veerman & Veldhuis-Diermanse, 2001; Pena-Shaff & Nicholls, 2004), the cognitive presence in community of inquiry (Garrison et al., 2001), the teaching of the community of inquiry (Anderson et al., 2001), and argumentative knowledge construction (Weinberger & Fischer, 2006). De Wever et al. (2006) compared 15 content analysis instruments from the perspective of the theoretical base, unit of analysis and inter-rater reliability and pointed out that existing analysis instruments need to be improved. Every content analysis scheme uses its own specific unit of analysis and data type. The analysis units are not identical in a variety of coding schemes, such as messages (Gunawardena et al., 1997), sentences (Fahy et al., 2001), paragraphs (Hara et al., 2000) and thematic units (Henri, 1992). The selection of the unit of analysis is very challenging for researchers. Although many researchers use “thematic unit” as the unit, the categorization standard of the “thematic unit” is very ambiguous. The complexity of interaction makes researchers use different vocabularies to code transcripts into different speech acts. For example, Fahy et al. (2001) coded transcripts into five kinds of speech acts (question, state, reflection, comment, and quote). The coding scheme developed by Pena-Shaff and Nicholls (2004) consisted of eleven kinds of speech acts (question, reply, clarification, interpretation, conflict, assertion, consensus building, judgment, reflection, support and other). Pilkington (2001) believes that coding schemes may categorize at too coarse a level to distinguish real communicative differences, or they may be too fine-grained to represent similarities. Porayska-Pomsta (2000) argues that categorizing speech acts is not useful in modeling teacher’s language and cannot account for the phenomena encounter in the dialogues. Furthermore, coding assigns each speech act an isolated meaning and does not record the indexicality of the meaning or contextual evidence (Suthers et al., 2010). In addition, the difficulty with content analyses of communications stems from a lack of guidelines for performing them validly and reliably (Rourke et al., 2001; Strijbos et al., 2006). Rourke et al. (2001) also discussed the importance of inter-rater reliability in the method of content analysis and pointed out that many researchers did not report coder reliability. Strijbos et al. (2006) believed that researchers should be cautious of the statistical test results when they did not report reliability parameters. The works of Dillenbourg (1999) and Stahl, Koschmann, and Suthers (2006) call for the need to develop process-oriented methodologies to analyze interactions. We believe that coding interaction transcripts into speech acts is very difficult because purposes of human’s speech acts are implicit; thus the identification of speech acts is very subjective. Simply focusing on the explicit speech acts will lead to ignorance of an individual’s knowledge construction. This study proposes an innovative method to analyze interactions in face-to-face collaborative learning. This method is IIS-map-based analysis method because it uses the IIS map. The whole study aims to validate the IIS-map-based analysis method that is used to analyze interactions and predict group performance. The empirical study is conducted to explore the effectiveness of the IISmap-based analysis method and to verify hypotheses. Methodology: IIS-map-based analysis method. Modeling and representing the collaborative learning system by IIS-map-based analysis method. You (1993) believes that the instructional system is a complex non-linear system that assumes cause and effect are associated disproportionately and the whole is not simply the sum of the properties of its parts. In addition, complex systems have an “emergence” property. Emergent properties arise at a particular level of system description by virtue of the interaction of relatively simple lower-level components - but cannot be explained at this lower level (Damper, 2000). Kapur et al. (2011) believes that the group is a complex system and convergence in group discussions is an emergent behavior arising from interactions between group members. Therefore the instructional system and collaborative learning system are both complex systems with characteristics of non-linearity and emergence. The complex systems cannot be understood by only analyzing visible factors such as teaching methods, various kinds of media, etc. To deeply understand various complex pedagogy phenomena and their effects, researchers should focus on the information flow within the system and its characteristics, as well as relationships between information flows and functions of the system (Yang & Zhang, 2009). We argue that the instructional system is an abstract information system. The collaborative learning system is a subsystem of instructional system, so it is also an information system. The function of a collaborative learning system is collaborative construction of knowledge by group members. Information processing and knowledge construction are closely intertwined in the learning process (Wang et al., 2011). The cognitive processes involved in knowledge construction are selecting relevant information from what is presented, organizing selected information into a coherent representation, and integrating presented information with existing knowledge (Mayer, 1996). The interconnection of the prior knowledge with the new information can result in reorganization of the cognitive structure, which creates meaning and constructs knowledge. Learning is a generative process of constructing meaning by linking existing knowledge and incoming information (Osborne & Wittrock, 1983). Based on the theoretical foundations, we argue that the nature of knowledge construction is to encode and decode information implicitly. Therefore information makes significant contributions to knowledge construction. Accordingly, the analytic focus is identified as information flows of the collaborative learning system. The information flow is defined as the output information of group members in the interaction process. The information flows between private information owned by each individual and the information shared by group members. In order to represent and analyze the collaborative learning system, a concept model is designed (see Figure 1). In this concept model IPL denotes the information processing of learners. IPL1, IPL2, IPL3, IPL4 denote information processing of multiple learners in one group. The internal information processes of IPL are not directly observable. However, the input and output information of IPL are visible. So {X} denotes the input information of IPL and {Y} denotes the output information of IPL. Because the output information {Y} is used for the purpose of sharing information, {Y} is abstractly generalized into an information set. This abstract information set is defined as Interactional Information Set (IIS). IIS is for sharing information in the interaction process. Thus {Y} is regarded as the input information of IIS and {X} is regarded as the output information of IIS. Vygotsky (1978) argue that learning takes place inter-subjectively through social interaction before it takes place intra-subjectively. IIS is generated and formed when information are externalized and shared in the social interaction process. Therefore IIS can account for social aspects of learning. We argue that IIS can represent the outcome of internal information processing of IPL. Because knowledge is constructed through processing information implicitly, some characteristics of IIS are closely related to the quality of co-construction of knowledge. The whole collaborative learning system is a functional coupling system which consists of IPL, {X}, {Y} and IIS. Figure 1. The concept model of the collaborative learning system. Coding and representing the input information of IIS. According to the concept model, three kinds of objects need to be represented. The first kind of object is the input information of IIS, namely {Y}. Because {X} is the input information of IPL and {Y} is the output information of IPL, {X} can be finally embodied and represented by {Y}, the analysis of {X} is unnecessary and the analysis focus is {Y}. In order to analyze the collaborative learning system, the attributes of input information of IIS need to be defined. These attributes include time, information processing of learners (IPLi), cognitive levels, information types, representation formats, knowledge network sub-map, annotation and the quality of information. Table 2 below shows the definition of each attribute. The coding format of input information items of IIS is defined as: [annotation] [the quality of information]. The symbol “< >” denotes the required item and “[ ]” denotes the optional item. Information flows in the interaction process will be represented and transformed into this kind of format. Table 2. The definition of each attribute. Visualizing information flows into sequences of IIS. In order to clearly represent information flows in the interaction process, the input information items of IIS are visualized into sequences according to the coding format. The sequences of input information items are the second kind of object that needs to be represented. Collaborators’ information items are chronologically on the timeline or under the timeline. The sequences of information items in face-to-face collaboration are linear sequences. Figure 2 shows the portion of sequential representation of IIS. The four long strands specify the portion of information sequences being shared at different time. Figure 2. The portion of sequential representation of IIS. The representation of IIS—— Knowledge Network Maps. The third kind of object need to be represented is IIS. We believe that IIS is closely related to the outcome of collaborative construction of knowledge. However, IIS is an abstract information set and it is not the real storage medium. Thus in order to represent IIS, a knowledge network map is used for visualizing knowledge and their relationships. From the beginning the knowledge network map is empty. As the interaction proceeds, the knowledge network map is gradually enlarged. Finally, it becomes the stable knowledge network map. The IIS-map is the knowledge network map with marks. The marks refer to attributes of information flows. The analysis of interactions can be reflected by the IIS-map. So this method is known as the IIS-map-based method. IIS-maps serve as abstract transcripts that can represent and visualize the interaction process. The Steps of IIS-map-based analysis method: An example. This section illustrates the steps of the IIS-map-based analysis method by using the example of co-constructing knowledge of graphs in the data structures curriculum. The IIS-map-based analysis method is conducted in three steps: Firstly, draw an initial IIS-map according to collaborative learning tasks. In this example, collaborative learning tasks are to design the algorithm of touring the campus so as to provide services for visitors. The learning objective of this task is to acquire storage structures of a graph, the algorithm of the minimum spanning tree and the shortest path of a graph. The detailed descriptions of tasks are as follows: 1. Design a campus plan, including the south gate, Jingshi Square, gymnasium, playground, library, technology building, and art building. How many kinds of storage structures can represent this problem? Explain the relative merits respectively. 2. Provide some information such as the name, code number and introduction of each scenery spots. 3. If sewer pipes are laid between each scenic spot, please design how to lay them down at the minimum cost. 4. If a visitor wants to tour from the Jingshi Square to the art building, please design the shortest path between them; In addition, please design the second shortest path between these two scenic spots. 5. If a visitor wants to tour the library and the technology building, please design the shortest path between the Jingshi Square and the art building, passing by the library and the technology building. 6. If a visitor sets out from the south gate and tours each scenic spot only once, and then leaves via the south gate, what will be the shortest tour path? The initial IIS-map can visually represent the domain knowledge structure. Generally speaking, the initial IIS-map represents teachers’ understanding of domain knowledge. The initial IIS-map is drawn according to the knowledge modeling norm (Yang, 2010a). This norm specifies seven categories of knowledge and eighteen kinds of relationships of knowledge. For more information refers to (Yang, 2010a). Figure 3 shows the portion of the initial IIS-map, in which the node represents the knowledge and the edge represents mutual relationship of the knowledge. Of course any kind of accepted norm can be adopted to draw the initial map. Figure 3. Portion of the initial IIS-map. Secondly, code and segment information flows into information items sequences of IIS according to the representation format. Specifically, the segmentation is in chronological order and it is based on the change of attributes of information items. The rules for segmentation and coding information flows are as follows: 1. When contributors of information change, the information flow will be segmented. 2. When cognitive levels change, the information flow will be segmented. The definition of cognitive level originates from the theory of instructional objective classification (Yang, 2010a). Discriminating, recalling and understanding are the same level. When the level changes into “applying” the information flow will be segmented because “applying” means a higher cognitive level. 3. When types of information change, the information flow will be segmented. 4. When the knowledge network sub-map changes, the information flow will be segmented. 5. The values of representation format do not influence the segmentation of information flows because there are many representation formats in each information flow such as text, sound, graphics and body language. All information flows in the interaction process need to be coded and segmented independently by two raters according to the rules. Table 3 shows fragments of information flows that originate from face-to-face interactions of a group. We code and segment information flows of Table 3 into sequences of information items of IIS, as shown in Figure 4. Table 3. Fragments of information flows. Figure 4. Portion of the sequences of input information items of IIS. Thirdly, compute attributes of information flows and generate the knowledge network map with marks. The marks refer to the quantity of activation. The quantity of activation is the information entropy that is generated by activating. We believe that some of the knowledge network maps are activated when information flows map the IIS. As long as the information flow has the knowledge network sub-map attribute, the knowledge is activated. The quantity of activation is an abstract attribute of information flows. It is designed to represent the quality of knowledge construction. Yang (2010b) proved that the quantity of activation of knowledge was positively correlated with learning outcomes in the context of classroom teaching. The higher the quantity of activation of knowledge is, the better learning outcomes will be. The algorithm of the quantity of activation of knowledge was designed by Yang (2010b) as formula (1): (FORMULA_1). Where: F is an adjustable parameter. F = 1 when there is <= 7 information items between the current information item and the preceding information item that activated the current knowledge. F = (k + 1) / (2k + 1) when there are more than seven information items. The value of k is equal to the number of information items between the current information item and the preceding information item that activated the current knowledge subtracting seven. Log (d+2) denotes the activation entropy and “d” is the number of the activated edges. Log (n*(D–d+2)) denotes the complexity. “n” denotes the categories of edges that are not activated. D denotes the total number of edges that connected with the vertex. r is an adjustable parameter. r = 1 when the vertex is directly activated. r =1/d when the vertex is activated only once by its adjacent vertex. r = 1/d2 when the vertex is activated twice by its adjacent vertex. The IIS-map-based analysis method aims to model and represent the collaborative learning system by information flows. It can visualize the interaction process and represent information flows onto the IIS map. The IIS-map-based analysis method regards the knowledge network map as the objective reliance body. The main content loaded by information flows is reflected and represented by the knowledge network map. This reduces the arbitrariness of coding discourse to a large extent, which makes the segmentation of information flows more objective and scientific. An Empirical study. This empirical study aims to verify if IIS-map-based analysis method can analyze interactions and some attributes of information flows can predict the group performance. The group performance is measured as the quality of knowledge construction, which can be calculated according to formula (2): (FORMULA_2). Where P denotes the difficulty coefficient of test items. The difficulty coefficient equals the ratio of the average score to the full mark of test items. CV denotes the coefficient of variation. It equals the percentage of the ratio of standard deviation to mean of score. Xiposttest and Xipretest respectively represent the score of pretest and posttest. N denotes the number of group members. This algorithm of group performance will be more objective in contrast to only computing the mean deviation between the pretest and posttest. The definitions of attributes of information flows. Interactions are very complex and some groups may deviate from learning objectives in the interaction process. Therefore on the IIS-map there is the targeting knowledge network map which is composed of the targeting knowledge. The targeting knowledge network can be selected and identified by teachers according to collaborative learning objectives. In this study we focus on one of the important attributes of information flows, namely characteristics of the targeting knowledge network map. Several indicators are designed to compute the characteristics of targeting knowledge network map, including the quantity of activation, average degree, average path length, density and degree entropy of the targeting knowledge network map. Quantity of activation of the targeting knowledge network map. Collaborative learning is different from the instruction which emphasizes the delivering of new knowledge to students. The precondition of collaborative learning is that the participants understand each other enough to accomplish their work (Stahl, 2011b). Furthermore, collaborative learning emphasizes collaborative construction of meaning among group members. Therefore the quantity of activation of the targeting knowledge network map is used for indicating the quality of knowledge co-construction. Many algorithms are tested. Finally, quantity of activation of the targeting knowledge network map is defined as the sum of quantity of activation of the targeting knowledge on the IIS map, as is shown in formula (3). (FORMULA_3). Where Ai denotes the quantity of activation of the targeting knowledge which can be calculated according to formula (1). N denotes the number of vertices of the targeting knowledge network map. Average degree. The average degree indicates the mean number of neighbors per vertex of the targeting knowledge network. The average degree can be calculated using formula (4): (FORMULA_4). Where E denotes the total number of edges. N denotes the total number of vertices. Average path length. The average path length indicates the scale of the targeting knowledge network. The average path length can be calculated using formula (5): (FORMULA_5). Where N denotes the total number of vertices. d ij denotes the quantity of edges of the shortest path between any two vertices. Density. Density can indicate the interconnectivity and closeness of associations of the targeting knowledge. The density of the vertex is calculated as the proportion of the number of actual edges to the number of possible edges. The density can be calculated using formula (6): (FORMULA_6). Where Ei denotes the number of actual edges. ki denotes the degree of the vertex. N denotes the total number of vertices. Degree entropy. The degree entropy shows the heterogeneity of targeting knowledge network. We adopt the algorithm of degree entropy in the complex network (Ferrer & Sole, 2003). It is defined as formula (7): (FORMULA_7). Where pk is the frequency of vertices having degree “k”. We can apply these measures to provide insight into the relationship between attributes of information flows and group performance. Hypotheses. We suppose that the following attributes of information flows can predict group performance. So five hypotheses of this study are proposed: H1: The quantity of activation of the targeting knowledge network map can predict group performance. H2: The average degree of the targeting knowledge network can predict group performance. H3: The average path length of the targeting knowledge network can predict group performance. H4: The density of the targeting knowledge network can predict group performance. H5: The degree entropy of the targeting knowledge network can predict group performance. As long as one hypothesis is verified, the effectiveness of the IIS-map-based analysis method is validated. Experiment Procedure. The experiment used pretest posttest research design. The section presents the step-by-step process for this study: 1. Setting collaborative learning objectives and selecting some knowledge as learning objects. In this study, tasks were originated from data structures. The detailed descriptions were shown in the section of the first step of the IIS-map-based analysis method. Teachers designed the collaborative task according to the learning objectives. This task was based on a real-life authentic situation, which was not explicitly taught in the course of data structures. Meanwhile, subjective testing items of the pretest and posttest were designed according to learning objective, which could be used to evaluate group performance. The detailed descriptions of pretest and posttest are in Appendix. The targeting domain knowledge included storage structures of a graph, the algorithm of the minimum spanning tree and the shortest path of a graph. 2. Drawing an initial IIS-map of the targeting domain knowledge. The initial IIS-map was composed of targeting knowledge and their relationships. The initial IIS-map was viewed as the standard map, and it represented what teachers expected to discuss in the interaction process (Figure 3). 3. Recruiting subjects and dividing them into different groups. The experiment selected thirty groups of subjects. There were three or four students in each group. The group members were divided into each group randomly. Then the subjects were placed in different laboratory rooms and collaborated via face to face. Group members collaborated for about two hours to solve the problem and designed at least one algorithm. Each group wrote down procedures of the algorithm in the last collaborative phase as the artifact. The pretest preceded the collaboration, followed immediately by the posttest. In order to record the authentic interaction process, we videoed the entire interaction process of thirty groups, and each group formed the total of approximately 120 minutes of video file in this experiment, according to which the initial knowledge network map was modified. In the experiment, the tasks were the same for the thirty groups, as were the questions for pretest and posttest. 4. Coding, segmenting and transforming information flows into sequences of information items according to the segmentation rules, as shown in Figure 4. 5. Computing the attributes of information flows, such as the quantity of activation, average degree, average path length, density and degree entropy of the targeting knowledge network, then generating the knowledge network map with marks, as shown in Figure 5. The numbers beside the knowledge in Figure 5 are the quantity of activation and they are calculated with formula (1). Finally analyzing and interpreting the relationships between group performance and attributes of information flows.   Figure 5. Portion of IIS-map with the quantity of activation. Figure 6. Screenshots of the analytic tool. Analytic Tool. We have developed the analytic tool that can draw the initial map, segment and transform information flows into sequences and compute attributes of information flows. Figure 6 shows the screen shots of the analytic tool. Inter-rater reliability. In this study, two raters coded and segmented information flows of thirty groups and assessed all test papers independently. One was the first author and the other was the research assistant, a graduate student majoring in computer science. The coders were trained on how to segment information flows and assess test items. After the training sessions, each rater coded information flows of thirty groups and assessed ninety-two test papers independently. In order to confirm the reliability of the segmentation and assessing the test items, the percent agreement statistic proposed by Holsti’s (1969) was used to evaluate for inter-rater reliability. The percent agreement index was most common used method in content analysis and it reflected the number of agreements per total number of coding decisions (Rourke & Anderson, 2001). The total checklist included all items of the pretest and posttest and segmented information items of thirty groups. The two coders discussed and resolved all discrepancies. Reliability coefficient was calculated and results showed all values were above 0.9, regarded as an indication of excellent agreement. Result. In order to test the hypotheses, the correlation analysis and linear regression analysis were conducted for group performance and attributes of information flows. Table 4 shows the mean and standard deviation of group performance, quantity of activation, average degree, average path length, density and degree entropy of the targeting knowledge network. Furthermore, the relationships between group performance and attributes of information flows were examined. Table4: Descriptive statistics for group performance and predictors. H1 assumed that the quantity of activation of the targeting knowledge network map could predict group performance. The result indicated that quantity of activation of the targeting knowledge network was significantly positively correlated with group performance (r = .559, p = .001). To examine the predictive validity of the quantity of activation on group performance, a linear regression analysis was conducted. The normal Q-Q plot was used to test normality of data. This test confirmed that the group performance variable had normal data. Consistent with hypothesis 1, the quantity of activation of the targeting knowledge network could predict group performance (t = 3.565, β = .559, p = .001). The quantity of activation could explain 28.8% of the total variance (Adjusted R2 =.288, F = 12.712). This indicates that the quantity of activation of the targeting knowledge network map is the significant predictor. The main reason is that the quantity of activation of the targeting knowledge network map can measure the semantic relationships and deep structure of knowledge network map. H2 assumed that the average degree of the targeting knowledge network could predict group performance. However, the data did not support this hypothesis: average degree was not correlated with group performance (r = .063, p = .739). So the average degree cannot predict group performance. The reason for this result is that average degree cannot represent the complexity of domain knowledge. H3 assumed that the average path length of the targeting knowledge network could predict group performance. However, the data did not support this hypothesis: the average path length was not correlated with group performance (r = .324, p = .081). So the average path length cannot predict group performance. The major reason is that the average degree is a simple indicator for the size of knowledge network map. H4 assumed that the density of the targeting knowledge network could predict group performance. However, the data did not support this hypothesis: the density was not correlated with group performance (r = .233, p =.215). So the density cannot predict group performance. The reason for this finding is that density cannot represent the deeper understanding of subject matter and semantic level. H5 assumed that the degree entropy of the targeting knowledge network could predict group performance. However, the data did not support this hypothesis: the degree entropy was not correlated with group performance (r = -.252, p = .180). So degree entropy cannot predict group performance. The reason for this is that degree entropy only indicates the static typology structure of knowledge network map. Discussion. The results of this study show that quantity of activation of the targeting knowledge network map can effectively predict group performance. This study also verifies that the IIS-map-based analysis method can analyze interactions in face-to-face collaborative learning. However, the results indicate that other four kinds of attributes including average degree, average path length, density and degree entropy of the targeting knowledge network cannot predict group performance. In fact these attributes can only reflect topological structures of the target"